Rectified Multi-class AdaBoost for Noisy Dataset Based on Weight Adjustment Standard

Abstract

Boosting, as a meta-algorithm for ensemble learning, have been widely applied to variety popular machine learning algorithms. However, noises in training and testing datasets could significantly affect the performance of boosting algorithm. SAMME pays too much attention to samples that are not correctly classified in multiple iterations. These samples could be mislabeled samples that cannot be correctly classified, so the classifier cannot learn the actual distribution of the original data. To solve this problem, in this paper, we proposed a rectified algorithm R.SAMME based on multi-class classification algorithm SAMME by limiting the weight of each sample based on current accuracy. We evaluate our approach on UCI benchmark datasets, experiments show that R.SAMME has better performance in noisy datasets.